CINS: Comprehensive Instruction for Few-Shot Learning in Task-Oriented Dialog Systems

As the labeling cost for different modules in task-oriented dialog (ToD) systems is high, a major challenge to learn tasks with least amount of labeled data. Recently, pre-trained language models (PLMs) have shown promising results few-shot learning ToD. To better utilize power PLMs, this paper proposes Comprehensive Instruction (CINS) that exploits PLMs extra task-specific instructions. We design schema (definition, constraint, prompt) instructions and their customized realizations three important downstream ToD, ie. intent classification, state tracking, natural generation. A sequence-to-sequence model (T5) adopted solve these unified framework. Extensive experiments are conducted on ToD realistic scenarios small validation Empirical demonstrate proposed CINS approach consistently improves techniques finetune raw input or short prompt.